Welcome to the hundreds of new subscribers who have joined over the past few weeks. In today’s free post, I take a look at what happens to Nvidia in a post-GPU future.
And if you like what you read here and you’re not yet subscribed, consider subscribing. Most of my posts are free though deeper dives tend to be paid.
Everyone’s still gorging on the Nvidia buffet, drunk on H100s and Blackwells like it’s 1999. Meanwhile, the real story is thermodynamics. Nvidia GPUs run hot. Hopper-class cards push 700 watts, and Blackwell is rumored to push 1000W per chip. This isn’t merely an engineering detail. It’s a systemic bottleneck.
AI’s scaling limits are not simply more flops or bigger models. They’re also about removing heat, and we’re slamming into a thermodynamic wall. Air cooling is maxed out. Liquid cooling is promising, but it is also expensive, complex, and still insufficient at scale. The entire stack is shaped by one brutal fact: you can’t cheat the laws of thermodynamics. Heat has to go somewhere.
And when cooler-running compute substrates emerge—photonics, analog in-memory computing, maybe even neuromorphic chips—the whole hardware ecosystem gets turned inside out. These post-GPU chips are cooler, and that changes everything.
Nvidia’s Moat is Thermodynamic
People talk about CUDA as Nvidia’s moat. That’s true, but only partially. The real moat is vertical integration across a heat-constrained stack. Nvidia sells chips, but they own the whole flow:
Chip design (Hopper, Blackwell)
Interconnects (NVLink)
Thermal solutions (liquid-cooled trays, rack design specs)
Data center tuning recommendations
Break the thermal envelope, and the whole integrated advantage begins to unravel. Cooler chips decouple software from heat. Suddenly, Nvidia’s core edge starts to matter a lot less.
And when that happens, competitors no longer need to catch up to Nvidia’s monolithic GPU engineering. They can route around it.
The Future is Horizontally Scaled, Thermally Liberated
Post-GPU architectures tend to be:
Distributed (modular, many small nodes)
Thermally cool (optical, analog, or mixed-signal compute)
Geographically flexible (edge inference becomes viable)
Non-monolithic (compute spreads out, not clusters in hotboxes)
Instead of centralizing compute into ever-hotter rooms, the future pushes it outward into cooler, simpler, cheaper hardware. If you remove the thermal constraint, why would you co-locate all your AI processing in Texas megastructures?
What Happens to Nvidia?
Let's assume photonic compute or analog in-memory chips start to show real performance per watt advantages by 2027-28. Nvidia can't buy these startups outright. Antitrust scrutiny after its failed ARM acquisition attempt makes that clear. But they can still pull off what Microsoft perfected in the '90s:
Control the abstraction layer.
Just like Microsoft used Win32 to control PC software regardless of who made the hardware, Nvidia uses CUDA and its associated stack (cuDNN, Triton, TensorRT) to own the developer mindshare and execution paths.
Does it matter if the chip is photonic or analog? Not if the compiler, runtime, and memory orchestration layer are all stamped with the Nvidia logo.
This is the playbook:
Acquire adjacent tech (e.g., optical interconnects like Ayar Labs) and frame it as system-level optimization.
Make strategic investments in next-gen chip startups, then tie them to exclusive SDKs and integration support.
Launch CUDA-light: Nvidia-branded developer tooling for third-party silicon.
Kill via bundling: Offer CUDA inferencing for free with cloud deals, undercutting independent chip upstarts.
Even if Nvidia loses the chip battle, it could win the compute war by abstracting itself above the hardware layer.
CUDA = Win32 for the AI Era
The best analogy here is Microsoft:
Windows was just the OS, but Win32 made it the platform.
Apps written for Win32 were sticky. Porting them to Linux or Mac was nontrivial.
Developers didn't leave, so users couldn't leave, so OEMs didn't leave.
Nvidia is now in the same position:
CUDA is the OS layer for AI developers.
Everyone stays because the tools are familiar, fast, and tightly coupled to hardware.
Nvidia doesn't need to win photonics or analog compute. It just needs to make sure those chips compile through Nvidia's stack.
The only way to break this grip is to go after the abstraction layer, not the silicon. But regulators are still playing whack-a-mole with hardware mergers. They're missing the real monopoly.
Conclusion: When the Heat Clears
Once you remove heat as the central design constraint, the entire topology of AI compute shifts. Nvidia's dominance is based on optimizing for an energy-dense, heat-saturated world. That world is ending.
If you're a data center operator, you're holding a melting asset.
If you're a cloud provider, you'd better hedge with modular, cooler-running form factors.
If you're a startup in photonics or analog compute, Nvidia will either smother you with love or shadow-ban you with CUDA incentives.
And if you're Nvidia, the smartest move is obvious: become the Microsoft of AI. Own the stack, not just the chip. Turn every emerging substrate into just another back-end target for CUDA.
Because in the end, the chip doesn't matter. The abstraction does.
Stay tuned for my next post about how a cooler compute future affects data centers.
Coda
If you enjoy this newsletter, consider sharing it with a colleague.
Most posts are public. Some are paywalled.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
follow me on Twitter,
connect with me on LinkedIn, or
send an email to dave [at] davefriedman dot co. (Not .com!)
Great piece, Dave. One near-term lever you didn’t mention: wide-band-gap silicon-carbide (SiC) power electronics.
• ~30-40 % of a rack’s load is overhead: AC-DC rectifiers, DC-DC converters, pumps, fans, UPS. Swapping legacy silicon switches for SiC lifts those stages from ~96 % to ~99 % efficiency, eliminating 70-100 W of waste heat for every 10 kW GPU node and cutting site PUE by ~8-12 % with <18-month payback.
• SiC ceramics as heat-spreaders/lids add ~2× thermal conductivity vs. today’s AlN, dropping GPU junction temps by 3-5 °C. Free head-room for those 1 kW Blackwells.
• It’s shipping (automotive volumes today, datacenter retrofit kits rolling out this year) so operators can buy time while photonics, analog in-memory, and neuromorphic chips mature.
SiC won’t replace Nvidia’s silicon logic, but it shrinks the thermal tax that threatens to stall AI build-outs right now- a practical bridge between overheated GPUs and the “post-GPU” future you outline.
Love the analysis.
In addition, it will be interesting to see how the cost curve changes. Right now I feel the software layer has little to no margin to make money as all the profits are with Nvidia. That has to shift a bit too. Sharing my article on the economics of it, primarily saddled by gpu costs
Tokenomics
https://open.substack.com/pub/pramodhmallipatna/p/the-token-economy
Private Model Econmics
https://open.substack.com/pub/pramodhmallipatna/p/private-model-economics-for-enterprise
Agent Economics
https://open.substack.com/pub/pramodhmallipatna/p/the-economics-of-ai-agents-a-high